Error in the `layers.conv` module

[mrzzmrzz]

Hello, I found a bug in the layers.conv module.

The function message_and_aggregate in almost all the models derived from the MessagePassingBase class is inconsistent with the function message and aggerate.

Take the RelationalGraphConv module as an example. In the message function, if there exists the edge feature, it first will be transformed by the self.edge_linear function:

 def message(self, graph, input):
        node_in = graph.edge_list[:, 0]
        message = input[node_in]
        if self.edge_linear:
            message += self.edge_linear(graph.edge_feature.float())
        return message

However, in the function message_and_aggregate, the edge_feature will be transformed according to its feature dimension change, which is inconsistent with the message and aggregate functions defined in the previous:

 if self.edge_linear:
    edge_input = graph.edge_feature.float()
    if self.edge_linear.in_features > self.edge_linear.out_features:
        edge_input = self.edge_linear(edge_input)
    edge_weight = edge_weight.unsqueeze(-1)
    edge_update = scatter_add(edge_input * edge_weight, node_out, dim=0,
                              dim_size=graph.num_node * graph.num_relation)
    if self.edge_linear.in_features <= self.edge_linear.out_features:
        edge_update = self.edge_linear(edge_update)

[Oxer11]

Thanks for reporting this bug!

This inconsistency is caused by the bias term in the edge_linear layer, which leads to the difference when applying before and after the scatter_add operation. I've fixed this bug in b41a764.